fix: decide constant folding structurally instead of by sampling (spec 07)

[minikin]

Summary

• GraphOptimizer::is_constant/is_constant_zero/is_constant_one previously decided constancy by sampling a node 3-4 times and comparing results — a low-entropy distribution (e.g. bernoulli(0.99)) had a high chance of sampling the same value repeatedly by chance and being silently folded into a constant, changing the distribution the optimized graph actually represents.
• ComputationNode::Leaf now carries constant_value: Option<T>, set only by the new ComputationNode::constant(value) constructor (and Uncertain::point, internally) or by the optimizer's own folded results. Every constancy check and constant-folding match arm reads this field directly — zero sample() calls anywhere in the decision path, matching the spec's own grep acceptance test literally.
• Discovered along the way (documented in the spec's implementation notes): GraphOptimizer isn't wired into Uncertain<T>'s arithmetic operators at all — it's an opt-in utility invoked manually. Unchanged by this spec, flagged for Spec 08 (effective CSE has the same characteristic).
• New acceptance-level tests: a deterministic low-entropy-distribution test (the bernoulli(0.99) scenario, now a hard guarantee rather than statistical), point(0.0)+x/point(2.0)*point(3.0) folding through the real public API, normal(0, 1e-12) never folded, and a same-seed optimized-vs-unoptimized equivalence test (sample_with + ChaCha8Rng).
• Updated examples/optimizations/graph_optimization.rs's literal 0/1 demo operands to use ComputationNode::constant — with the old sampling-based check removed, the demo would otherwise silently stop demonstrating identity elimination/constant folding while still printing the same correct final number.
• Not a breaking change — public API only gains ComputationNode::constant; Uncertain::point's signature/behavior is unchanged.

Test plan

• just dev green (fmt, clippy -D warnings, tests default + parallel, doc tests, audit, deny, crap regression gate)
• Full test suite + doctests pass with --all-features and --features parallel
• Verified via awk/grep that no sample() call appears in is_constant_zero/is_constant_one's bodies
• Ran examples/optimizations/graph_optimization.rs manually — still prints the correct 32.0 result both optimized and unoptimized
• crap_baseline.json regenerated (10 functions improved from simplified call sites, 2 now-unused helpers removed)
• CHANGELOG.md/specs/README.md updated